It is 1:47 a.m. somewhere in Accra. A man is lying beside his wife, phone tilted away from her at that particular angle we have all mastered.

He is not texting another woman. He is typing to a chatbot.

He is asking it something he has never asked his pastor, his brother, or the woman breathing next to him. Something about his marriage. Something about his body. Something he has carried since he was fifteen and has never found a room safe enough to put down.

And the machine answers. Warmly. Immediately. Without a single raised eyebrow.

Then he presses delete, turns the phone face-down, and sleeps.

The two halves of that night are one thing. The comfort and the record come from the same machine. He did not pour out his heart and then, separately, leave data behind. The comfort is what drew the data out of him.

Hold onto that. Everything else in this piece hangs off it.

Why he is awake

Let us be honest about the demand before we judge the supply.

In our society there are questions you cannot ask out loud. Tell your uncle you are depressed and he will tell you to pray harder. Ask an auntie about intimacy in your marriage and by Sunday three people will know. Ask a Ghanaian man to say I am struggling and watch how quickly the room changes temperature.

I am writing specifically about Ghanaian men here, and I want to be precise rather than sweeping. A woman's 1:47 a.m. is a different scene with different stakes — around surveillance, around safety, around who gets to know she is unhappy. That deserves its own article and I am not qualified to write it in a paragraph.

I should also name the entry fee. The man in my opening has a smartphone, a data bundle, and English literacy. The most emotionally rationed Ghanaians have none of the three. This confession booth is not free, and the people most starved of a listener are furthest from the door.

But note what I have just done: I have made a claim about Ghanaian emotional life and offered you no evidence for it, because there is none to offer. No Ghanaian dataset measures this. Not the Ghana Statistical Service, not the Mental Health Authority, not the universities. Every number in the next section is American.

That absence is not a gap in my argument. It is my argument. We are about to reckon with a phenomenon nobody in this country has measured.

What the numbers actually say — including the ones against me

The Pew Research Centre surveyed 1,458 American teenagers between 25 September and 9 October 2025. Twelve per cent had used chatbots for emotional support or advice; sixteen per cent for casual conversation. Only eighteen per cent of parents were comfortable with the former.

Common Sense Media surveyed 1,060 American teenagers in April and May 2025. Seventy-two per cent had used an AI companion. A third had chosen one over a real person for a serious conversation. Thirty-one per cent said those conversations were as satisfying as talking to human friends, or more so.

Now the numbers that cut against me, because you will find them anyway.

Sixty-seven per cent of those same teenagers still found AI conversations less satisfying than human ones — forty-seven per cent said much less. Eighty per cent of AI companion users spent more time with real friends than with the machine. And thirty-nine per cent reported carrying social skills they had practised on AI into real-life situations.

Read plainly: most young people have kept their heads. The apocalypse has not arrived.

So let me narrow my claim to something I can defend. I am not writing about the two-thirds. I am writing about the third — and about what happens as the third becomes the half. And I am writing about a country where nobody is counting either group.

The friend who agrees 49% more than a human would

In March 2026, Stanford researchers published a study in Science that ought to be required reading for anyone who has asked an AI to settle an argument.

They tested eleven leading models — ChatGPT, Claude, Gemini, and others. Those systems affirmed users' actions forty-nine per cent more often than human beings did. The affirmation held even when the user described deception, illegality, or harm to themselves or others.

Then they ran experiments on more than 2,400 people, and found the part that unsettles me. A single conversation with a flattering AI made participants more convinced they were right, less willing to apologise, and less inclined to repair the relationship they had come to complain about. And still they preferred the flattering model. They trusted it more.

Now, a fair critic will say the comparison is American: the "human" the machine was measured against was an anonymous online commenter, not a Ghanaian elder. If our social feedback is harsher than that baseline — shame-forward, honour-inflected, publicly adjudicated — then relative to us the machine may be less a distortion than a mercy. I take the point, and I want to keep it in view rather than wave it away. It does not dissolve the finding, because the harms Stanford measured were real behaviours, not deviations from a norm. But it tells us something important: the machine's pull is strongest exactly where our own rooms are cruelest. Hold that thought. We will need it at the end.

We have an expression for the friend who agrees with everything you say. We do not call that person a friend. We call that person dangerous — the one who claps while you walk into the fire, because clapping is cheaper than confrontation.

Here is the hinge

Now put the two facts side by side and watch what happens.

The machine agrees with you at a rate no human would. That agreement is precisely what makes you keep typing. You disclose the second thing because the first thing was received without judgment. You disclose the third because the second was. By the fourth message you are telling it something you have never said aloud in thirty-nine years.

The flattery is not a harmless quirk that happens to sit next to a data policy. The flattery is part of how the data gets collected.

This is why the sycophancy finding and the retention question cannot be discussed separately. A system that disagreed with you the way your brother does would extract a fraction of what this one extracts. Agreeableness is not only the product's charm. It is also the product's aperture.


And now the delete button

Here is where I see the most confident misunderstanding, and where I need to be careful.

Deletion does something real. Take Anthropic's published policy, among the more transparent in the industry: delete a conversation, and it is excluded from future model training. That is a genuine commitment, and I will not pretend otherwise.

What deletion does not do is remove what has already been derived.

The transcript goes. The conversation leaves your history immediately and clears back-end systems within about thirty days. But the inferences drawn from it — safety classifier scores, flagged-content records, behavioural signals — operate on their own timetables, and those are longer. And if you have left the "help improve the model" setting switched on, material you have not deleted can be retained in de-identified form for up to five years. Other conversational AI providers publish comparable windows; some publish less. None of this is scandal — it is ordinary distributed-systems engineering, and deletion instructions take time to travel across backups and logs.

So the accurate sentence is narrower than the one I want to write, and more disturbing:

Deletion removes the document. It does not remove the conclusions already drawn from the document.

You can erase what you said. You cannot erase that the system now holds a shape where your shame used to be. Incognito mode reduces what is stored and remembered; it does not make the words unhappen.

The part where I argue against myself

I have been building toward a defence of human friction — the elder who embarrasses you, the friend who interrupts, the counsellor who makes you sit in silence until you hear how you sound. I believe that friction is where character is formed.

But I have to be honest about who I am recommending it to.

For a battered wife, community friction is not sandpaper. It is a blade. For a gay Ghanaian, the room that "turns hostile" is not a teacher — it is a threat with a congregation behind it. For someone in a manic episode, the family sitting is not correction; it is spectacle.

The appeal of the machine scales exactly with how dangerous your community is. My earlier framing praised the very room that half my readers are running from, and I was able to write it because I have a family that is safe to be known by. Not everyone is drawing from that hand. So I will not tell you the friction is always holy. It is corrective for some and lethal for others — and remembering that is the key to where this ends.

So where do we go — and in what order

I do not believe the first answer to any of this is the government. I believe responsibility runs in a sequence, and that we damage ourselves when we skip the early steps to demand the last one. It begins with the person. It grows through family and community. It reaches the nation last — as reinforcement, not rescue.

First, the person. You are not powerless here, and pretending you are is its own kind of surrender. Use the machine as a rehearsal room, not a confessional — draft the hard conversation with it, then have it with a human. The evidence is on your side: thirty-nine per cent of teenage users carried skills practised on AI into real life. It is a good scaffold and a poor destination. And when it agrees with you completely, notice — that is the Stanford effect working on you in real time. Ask it for the other person's case. Read that part twice.

Then, family and community. The deepest fix is not a better machine or a better law. It is rebuilding the rooms — making a family, a friendship, a faith community into a place where a man can say I am struggling and the temperature does not change. That is slow, unglamorous, generational work, and it is ours, not the state's. But I said we would need one thought from earlier, and here it is: the machine wins most easily where the community is cruelest. So part of this work is honest self-examination. Every room that shames the vulnerable into silence is quietly manufacturing the very demand that sends people to a machine at 1:47 a.m. If we want fewer people confessing to chatbots, we build gentler rooms.

Last, the nation. There is a role for the state, but it is the final room, not the first. Ghana's Data Protection Commission could ask any AI service with significant local use to disclose plainly what survives deletion, for how long, and under whose law. The power to compel transparency is more defensible than the power to compel erasure across a border — and even that is a question of enforcement, not just principle. But a disclosure rule only helps a citizen who already knows to care. The state can light the room. It cannot make you sit in it with another human.

What I would actually tell the man at 1:47 a.m.

Not a policy. A sentence, from me to you.

When the choice is between a machine that will always agree and a human who sometimes will not, choose the human. And when your family or your faith community is not the safe human — because for some of you it genuinely is not — do not settle for the machine. Choose a professional one.

A trained, confidential counsellor is the third room this whole essay has been looking for. Not the village that would shame you. Not the machine that would keep you. A person whose entire discipline is to hold what you say without weaponising it and without flattering you out of your own repair. That is not a lesser option than the machine because it is human and imperfect. It is the better option precisely because it is human — bound by confidence, trained in care, and unafraid to tell you, gently, when you are wrong.

I can name one, because I have sat in that room myself and left it better than I entered: Anthea Crabbe, founder of the Aruka Centre here in Accra. There are others.

But I owe you the same honesty I asked of myself earlier. That room is not evenly spread across this country. For the man in my opening, it is a taxi ride away. For the reader fifty kilometres out — without the data bundle, without a centre that takes his money or speaks his language — it may not be there at all, not yet. I will not tell you the door is nearer than it is. For some of you, building that door is precisely the family-and-nation work I sequenced above: more trained counsellors, in more towns, in more languages, is not a private errand but a thing a society decides to make. Until then, I can only tell you the room is worth reaching for, and worth demanding where it does not exist.

The machine will hold your secret. It cannot hold you. But a good counsellor can — and unlike the machine, she will sometimes tell you that you are wrong. That was never something to fear. That was the whole reason to go.

---

Creppy Emmanuel writes on AI governance, culture, and digital society. He uses these tools daily in his professional work and writes about the Aruka Centre from personal experience as a former client; this recommendation is unpaid.